AI · ML · GENAI ENGINEER INTERVIEWS / 957 QUESTIONS

Ace every round of the AI interview.

Know what each round is testing, before you are in it.

The AI engineer interview is a loop of rounds, and each one screens for something specific. This site maps 957 real questions, 230 concepts and 40 company loops onto that sequence, so you prepare for the round you are actually about to sit, not a generic one.

Start free →See the loop ↓

110 answers free, no account · Premium $25 or ₹2,000 once · 6 months · no subscription

957 questions answered in full230 concepts with self-check drills40 company loops · 25 documented round by round110 answers free without an accountGrounded in real GenAI, LLM, and AI/ML engineering interview loops and written to a senior-engineer editorial bar.

THE LOOP

Five rounds. Every one of them has a question bank, and every sample here is free to read.

Loops differ by company, but the rounds recur. Each one below shows the topics it draws on, with live counts, and two questions the way they are actually asked.

R1
Recruiter and hiring-manager screen
Whether your story matches the role, and whether you know which loop you are walking into.
Behavioral & Project Deep-Dives62 What the role actually is The loop at your target company
62 questions in this round
- Tell me about a time you had to learn a new technology or domain quickly, and how you keep up with AI.mediumBehavioral & Project Deep-Divesfree
- Walk me through a project you're most proud of, in depth.mediumBehavioral & Project Deep-Divesfree
R2
Coding
Practical builds and the LeetCode-medium staples, with the data plumbing underneath.
Coding & DSA144 SQL & Data Engineering58
202 questions in this round
- Given a query vector and N stored vectors, return the top-k most similar by cosine similarity, efficiently.mediumCoding & DSAfree
- When do you choose batch vs streaming, and what are the Lambda and Kappa architectures?mediumSQL & Data Engineeringfree
R3
ML and LLM fundamentals
The conceptual core: attention, sampling, evals, metrics, calibration, the modeling basics still tested beside GenAI.
LLM & GenAI Fundamentals135 Machine Learning & Data Science126 The concept curriculum Every term, defined
261 questions in this round
- Define precision, recall, F1, and AUC, and give a case where each (and accuracy) is misleading.mediumMachine Learning & Data Sciencefree
- How do you tell whether model A is genuinely better than model B, not just better by chance?mediumMachine Learning & Data Sciencefree
R4
System and ML design
The modal round: RAG and agents, production trade-offs, GPUs, MLOps, security, and the model behind a product.
RAG & Agent System Design104 System Design for AI in Production119 ML System Design (Product)12 ML Infrastructure & GPUs70 MLOps & ML Engineering60 AI Security, Privacy & Governance67 Agentic AI, specifically
432 questions in this round
- Design a text-to-SQL feature: let users ask questions in natural language over a real database.hardSystem Design for AI in Productionfree
- Design an evaluation framework for an ads-ranking system.hardML System Design (Product)free
R5
Behavioural and project deep-dive
The highest-variance rounds and the least prepared for: the system you shipped, the model that failed, the trade-off you owned.
Behavioral & Project Deep-Dives62
62 questions in this round
- Tell me about a time you received critical feedback or made a mistake. How did you handle it?mediumBehavioral & Project Deep-Divesfree
- Tell me about a time you drove a decision or change across teams without having authority over them.mediumBehavioral & Project Deep-Divesfree
↻
Then the loop at the company that is actually interviewing you.
40 company guides. 25 loops documented round by round with what each stage screens for; 10 partly verified and marked as such. Where we are not sure, the page says so.
OpenAI Anthropic Google Google DeepMind Microsoft AWS xAI Salesforce Hugging Face Scale AI Shield AI SpaceX ByteDance Palantir Meta Databricks Snowflake Cohere Mistral Glean Sierra Harvey NVIDIA Perplexity C3 AI all 40 →

WHAT AN ANSWER LOOKS LIKE

Not a list of questions. Answers with the parts an interviewer is listening for.

A question you can look up anywhere. What a room rewards is the shape of the answer: what to say first, where the depth is, the follow-up they hold in reserve, and the mistake that sounds right. Every one of the 957answers here carries all four. This is today's, unedited, from the free window.

Grounded in real GenAI, LLM, and AI/ML engineering interview loops and written to a senior-engineer editorial bar.

TODAY'S OPEN ANSWERSQL & Data Engineering · hard

Design a data pipeline that is safe to re-run: idempotent writes, late data, and exactly-once effects.

TL;DRwhat you say if they cut you off: Exactly-once delivery is not something you can guarantee, so engineer exactly-once effects through idempotency: turn every write into a deterministic upsert keyed on a stable business/event key (MERGE or insert-overwrite-by-partition), so reprocessing identical data lands on the same result. Deal with late data using event-time windows plus a watermark that caps how long you wait, and checkpoint progress so a retry p…
HOW TO APPROACH ITthe order, and the clarifying questions: Fix the framing up front: distributed systems deliver at-least-once, so the aim is idempotent processing, not some magical exactly-once. Then walk the three mechanisms in sequence: idempotent writes, partition/window scoping, and watermarking for late events.
THE FOLLOW-UPwhat they probe next, with the strong reply: "Why not rely on exactly-once delivery?" It is fragile and usually holds only within one system; idempotent writes survive retries, replays, and backfills regardless.
THE MISTAKEwhat sounds right and loses the offer: Blind INSERT/append, so a retry double-counts. Processing on wall-clock time, so late events land in the wrong window or are lost. No watermark, so windowed state grows unbounded. Advancing the source offset before the write commits, losing data on a crash between the two.

Read the whole answer →with the code, the diagram and the discussion

AFTER TWENTY QUESTIONS

It starts telling you where you are weak.

Reading answers is passive, and a chat window forgets you when you close it. One tap after each answer, Got it, Shaky or Review, feeds a per-topic recall estimate built on knowledge tracing, the model behind intelligent tutoring systems. Your next focus is your weakest round, not a random question.

free with an account · your progress persists between sessions

# example session log
session   4 · rated 18 · 6 topics touched
system-design   ██████████░░  strong 78%
rag-agents      ████████░░░░  strong 61%
llm-genai       ██████░░░░░░  developing 44%
mlops           ███░░░░░░░░░  building 22%

next focus  → mlops · Design a retraining trigger from drift, not a schedule

PREMIUM

Try the depth first. Then one payment for 6 months of all of it.

✓All 957 answers. 110 are open to anyone, 212 to a free account, and 3 premium answers are free to taste. Premium is the other 745.
✓All 230 concepts and every self-check drill.
✓Mastery tracking that names your weakest round and what to do about it.
✓Everything added during your 6 months, at no extra cost.

$25~~$50~~ launch price

or ₹2,000 ~~₹4,000~~ · one payment · 6 months

Less than an hour with an interview coach, for the whole job hunt. Card, UPI or PayPal. Nothing recurring, no card stored, nothing to cancel. Access unlocks the instant payment confirms.

Unlock everything →Or keep reading free: 10 answers per topic, 20 with an account

ELSEWHERE ON THE SITE

Ace every round of the AI interview.

Five rounds. Every one of them has a question bank, and every sample here is free to read.

Recruiter and hiring-manager screen

Coding

ML and LLM fundamentals

System and ML design

Behavioural and project deep-dive

Then the loop at the company that is actually interviewing you.

Not a list of questions. Answers with the parts an interviewer is listening for.

Design a data pipeline that is safe to re-run: idempotent writes, late data, and exactly-once effects.

It starts telling you where you are weak.

Try the depth first. Then one payment for 6 months of all of it.